Biomedical data, which encompasses EHRs, medical imaging, sensor outputs, omics, and unstructured text, is becoming an ever-greater obstacle to healthcare reform due to its complexity and heterogeneity. There is a lack of fully realized data-driven medical insight potential due to the fact that these data types are frequently multi-dimensional, badly annotated, and hard to analyze. Aiming to circumvent this, precision medicine integrates environmental elements, lifestyle variables, genetic variability, and electronic health records (EHRs) to guarantee individualized treatment for individuals. This research proposes a new classification method that combines BERT with LightGBM to solve these problems in AI healthcare. The approach uses feature selection methods to keep the most important properties and thorough data pre-processing to guarantee consistent and noise-free input. Using AI Healthcare datasets, the suggested BERT-LightGBM model was compared to four other classification models that used different word embedding methodologies in order to assess performance. Our proposed model outperformed the competitors with a precision of 93.32% and an accuracy of 95.43%, as shown by the results. These results show that deep contextual language models combined with gradient boosting are a powerful tool for AI healthcare advancement and precision medicine prediction accuracy improvement. © 2025 IEEE.